← All guides

How to Write a Prompt for an AI Voice Agent

Why the prompt is the product

Two businesses can buy the same voice agent platform and get wildly different results, and the difference is almost never the software. It is the prompt: the written instructions that define who the agent is, what it may say, what it must never invent, and when to hand over to a human. Voice agents are literal-minded employees with perfect memory and no common sense, which makes them exactly as good as their briefing. The good news is that a receptionist prompt is not code and not art — it is six sections of plain writing, most of which you already know about your own business. This guide walks through each section, and the builder linked below turns your answers into a finished prompt automatically.

Section 1 — Role and tone

Open by telling the agent who it is: 'You are the receptionist for Rapid Flow Plumbing, a family-run plumbing company in Austin.' Add the tone in a few adjectives a human would understand — friendly and brisk, warm and unhurried — and any identity rules: if asked, the agent politely confirms it is an AI assistant and offers a human callback. That disclosure line matters legally in some jurisdictions and ethically in most; it also disarms the callers who would otherwise spend the call testing whether they have been fooled. Keep the role narrow. An agent told to 'handle everything' will try to, including the things it should never touch; an agent with a job description behaves like it has one.

Section 2 — Greeting

Write the greeting out word for word, because improvisation at second zero sets the whole call. Keep it under fifteen seconds, business name first: 'Thanks for calling Rapid Flow Plumbing — how can I help you today?' If most out-of-hours calls are emergencies, your platform may support a different greeting at night that acknowledges urgency; if not, the single greeting plus fast triage still works. Avoid greetings that advertise the technology, promise things the agent cannot deliver, or list options like a phone menu — the entire point of a voice agent is that the caller can just talk, so let them, immediately.

Section 3 — Tasks, in order

List the agent's jobs as numbered tasks in the order they occur: identify the caller's need; book, reschedule or cancel using the connected calendar; answer questions using only the knowledge section; take a complete message otherwise. For each task, specify the data that makes it complete — a booking is not done until name, number, address, job description and time window are confirmed and repeated back. Vagueness here is what produces agents that cheerfully 'book you in' with no address. If your platform supports structured steps or workflows, the prompt still matters: it governs everything between the steps, which is where calls actually live.

Section 4 — Knowledge, and the rule about inventing

Give the agent a closed world of facts: services, service area, hours, published prices and fees, policies, parking, payment methods — and then add the most important sentence in the whole prompt: if you do not know, say so and offer a callback; never guess a price or a policy. Language models would rather invent a plausible answer than admit a gap unless instructed otherwise, and an invented price quoted to your customer is your price in their mind, whatever your invoice later says. Write knowledge entries in the caller's vocabulary, not your internal jargon, and update the section whenever your facts change — an out-of-date prompt is a staff member confidently working from last year's price list.

Section 5 — Escalation

Define the exits. Escalation triggers typically include emergencies specific to your trade, distressed or angry callers, requests for a named person, anything outside the knowledge section, and any caller the agent has misunderstood twice. Specify what escalation means in your setup: a live transfer to the on-call phone, or an urgent message plus a promised callback time — and have the agent say which is happening, plainly. Callers accept 'this needs our on-call plumber, who will ring you within fifteen minutes' remarkably well. What they do not accept is being looped, deflected or smoothed over while their problem grows. The escalation section is where your agent's humility lives, and humble agents get forgiven their mistakes.

Section 6 — Closing, and then: test

Close every call the same way: summarise what will happen next and by when, confirm the callback number, thank the caller by business name. It takes ten seconds and prevents the most common post-call failure — a customer who is not sure anything was actually booked. Then test the whole prompt the way an awkward customer would: an out-of-hours emergency, a price you deliberately left out of the knowledge, a request outside your services, a caller who mumbles. Fix failures in the text of the relevant section, not by adding exhortations to 'be careful'. When your test calls bore you, the prompt is done — and if you would rather not start from a blank page, the builder below assembles all six sections from a short form.